First Insights into Hybrid AI-Fuzzy Tutoring System

for Boredom Identiﬁcation

Viktors Zagorskis

1 a

, Ingrida Lavrinovica

2 b

and Atis Kapenieks

Distance Education Study Centre, Riga Technical University, Kronvalda 1, Riga, Latvia

Telecommunications Institute, Riga Technical University, Azenes 12, Riga, Latvia

Keywords:

e-Leaning Quality, Computer Agent, Artiﬁcial Intelligence, Emotions Recognition, Fuzzy Intelligent Control.

Abstract:

In this paper, we introduce the Hybrid AI-Fuzzy Intelligent Control (HAFiC) system prototype. The proposed

model is the add-on to online learning-tutoring environments to proactively detect learners’ emotional states

by measuring performance gaining or degradation in a game-like form. We introduce a system model and

experimental results implementing recently proposed Simple Algorithm for Boredom Identiﬁcation (SABI)

(Zagorskis et al.,2019) along with the Fuzzy Intelligent control approach to evaluate whether proposed in-

direct data acquisition method allows retrieving performance data variability in correspondence to real user

emotional states. In the proposed system, the AI part cares about Image Processing and Text Recognition gath-

ered from mobile-handwriting devices. In contrast, Fuzzy Expert System part organises users’ performance

data utilisation and decision making based on adaptive fuzzy inference approach. First experiment results

described.

1 INTRODUCTION

This paper introduces with experience of a building

of hybrid intelligent learning-tutoring system mock-

up powered by Artiﬁcial Intelligence (AI), Machine

Learning (ML) (Russell and Norvig, 2016), and

Fuzzy Logic (Siddique, 2013) techniques.

Learning environments centered on the student

can help to address some weaknesses of traditional

education models such as complexity in learner’s per-

formance management and generation of personal-

ized learning environments (Lugo et al., 2015). Re-

spectively, there is a need in solutions that would meet

the learners’ needs, interests, rhythms and styles. Ac-

cording to the latest trends in e-learning education

market, personalized content is presented as an in-

telligent machine response to learner’s requirements.

Therefore, design of a personalized learning environ-

ment makes it necessary to cover the following set of

characteristics: understanding the situation of the stu-

dent in terms of emotional and cognitive states, previ-

ous knowledge background, skills, interests, response

to situations related to the teaching learning process

and learning style (Lugo et al., 2015). Individual

https://orcid.org/0000-0002-6155-0570

https://orcid.org/0000-0002-0343-167X

learning style directly characterizes student’s behav-

ior and plays a key-role in building predictive model

according to cognitive state analysis. For example,

sensing students prefer facts, data and experimenta-

tion whereas intuitive students prefer principles and

theories. Sensing students are patient with detail but

do not like complications, whereas intuitive students

are bored with detail and welcome complications (Li

and Rahman, 2018).

In this paper, we give insights into the implemen-

tation of Simple Algorithm for Boredom Identiﬁca-

tion (SABI) (Zagorskis et al., 2019) in the experimen-

tal mockup. The essence of the SABI algorithm is

boredom identiﬁcation analyzing periodic handwrit-

ing on mobile surfaces. Afterwards, we validate SABI

algorithm operation in action.

Aim of the research is to evaluate whether indi-

rect data acquisition method realized through SABI

algorithm, AI-based and Fuzzy Intelligent control

methods allows getting data variability suitable for

a successful boredom detection and optimization of

e-learning ecosystem architecture and technology re-

spectively. Hypothesis is formulated based on the re-

sults from previous research (Zagorskis et al., 2019).

The paper has been organized as follows: Sec-

tion 2 contains the reﬂection of related theories, meth-

ods, and approaches; in Section 3, which is the most

320

Zagorskis, V., Lavrinovica, I. and Kapenieks, A.

First Insights into Hybrid AI-Fuzzy Tutoring System for Boredom Identiﬁcation.

DOI: 10.5220/0009413203200326

In Proceedings of the 12th International Conference on Computer Supported Education (CSEDU 2020) - Volume 2, pages 320-326

ISBN: 978-989-758-417-6

important part of the research, we propose design of

the hybrid Fuzzy-logic based tutoring system; in Sec-

tion 4, we describe the experimental results; Section

5 concludes the paper.

2 PRELIMINARIES

2.1 Identifying Learners’ Activity

Each learner’s individual characteristics like motiva-

tion, attitude, learning style, background knowledge,

ability to apply preliminary skills and the number of

asked questions - affect the learning process (Graf

et al., 2010). In the e-learning environment, know-

ing speciﬁc characteristics of user activity, it is pos-

sible to reveal hidden personality features like a ten-

dency to be bored or being engaged in the learning

process. Compared with other emotions, such as anx-

iety, boredom can be considered a relatively ’silent’

emotion (Tze et al., 2016). Summing up the facts

above, our research covers an evaluation of the re-

lationship between student’s academic boredom and

potential learning outcomes.

Analysis of user’s activity data will allow us to ap-

ply AI or ML methods to identify the user’s learning

style. Moreover, by conducting user self-assessment

and emotional states surveys, we will get one more

data tensor for analysis and supervising of machine

methods. Comparing both data, we can make con-

clusions regarding user behaviour patterns, emotional

conditions and learning styles.

Nowadays, instructional designers follow state-of-

the-art practices to engage learners by new instruc-

tional design and content. It has been noted that even

the same content enveloped in different frames can

have a disparate impact on learners attention time.

Learners’ activity and engagement deﬁned by inter-

action environment settings gains better user experi-

ence (UX) and instructional design in a game-like ap-

proach.

2.2 Recent Boredom Detection Methods

There are at least two methods for boredom detec-

tion: direct (or explicit) and implicit method. Ex-

plicit implementation of surveys (or interviewing) is a

kind of intelligent methods involving post-processing

of the textual, audio, and video data. It is applicable

for the usage in small-scale environments while for

large-scaled online learning environments automated

and implicit methods are more relevant.

Environments. In recent years boredom detection

based on automatically inferring user behaviour activ-

ities from mobile phone usage has been actively stud-

ied. M.Pielot et al. introduce the user-independent

machine-learning model of boredom-leveraging fea-

tures related to (1) recency of communication, (2) us-

age intensity, (3) time of a day, and demographics

(Pielot et al., 2015). As results show, M.Pielot et al.

’s experiments infer boredom with an accuracy of up

to 82.9%. Before that, Bixler and D’Mello show that

the most popular methods for boredom detection are:

(1) facial expressions, (2) speech and voice features,

text analysis, and physiological signals from mobile

devices (Bixler and D’Mello, 2013).

Self-Assessment Manikin (SAM) method

(Bradley and Lang, 1994) evaluates the pleasure,

arousal and dominance in one scale. In SAM method,

the perceived users’ emotions are obtained by ques-

tionnaires and recorded in three-dimensional spaces

(valence, arousal, and dominance). Each space has

ﬁve ranking levels giving the capability to identify

15 different emotions. Although the boredom is not

given in the emotions list, the new methods can be

elaborated modifying SAM method. For example,

fuzzy Hidden Markov Chains (FHMC) we ﬁnd as one

of possible research directions in the future (Wang

et al., 2014), (Cannarile et al., 2018).

Stochastic Multi-player Games (SMGs). Another

way of modelling stochastic processes is the usage

of games-based methods. For example, for stochas-

tic two-player game modelling two competing entities

can be used: a) a user working on it’s mobile device,

and b) machine or host operating according to its own

or general strategy.

To specify the system evolution by the decisions

of multiple players taking into account the presence

of their probabilistic behaviour, we constrain our at-

tention to turn-based to stochastic games, in which a

single player controls

Deﬁnition One. (Stochastic Multi-player Game)

A stochastic multi-player game (SMG) is a tuple G :=

(Π, S, (S

)

i∈Π

, s, A, δ, L), where:

• Π is a ﬁnite set of players,

• S is a ﬁnite state of states,

• (S

)

i∈Π

is a partition of S,

• ¯s ∈ S is an initial state,

• A is a ﬁnite state of actions,

• δ : S × A → Dist(S) is a probabilistic transition

function,

• L : S → 2

is a labeling function mapping states

to sets of atomic propositions.

First Insights into Hybrid AI-Fuzzy Tutoring System for Boredom Identiﬁcation

321

Probabilistic behaviour typically coexists with

non-determinism. Both non-determinism and prob-

ability (in discrete state space) are present in

the classical model of Markov decision processes

(MDPs), which typically expressed in temporal log-

ics (Kwiatkowska, 2013).

Markov decision processes (MDPs) represent the

case when SMG contains only one player - Π := {1}.

Most common and widely studied class of SMG mod-

els involve two-players (Π := {1, 2}).

Deﬁnition Two. (Strategy) A strategy for player i ∈ Π

in SMG G is a function σ

: (SA) · S

→ Dist(A).

Summing up the above facts, we formulate the fol-

lowing research question: RQ - does indirect data ac-

quisition method allow getting data variability related

to the SABI algorithm?

About Intelligent Control Methods. Notably that

intelligent control systems are not deﬁned in terms

of speciﬁc algorithms (Siddique, 2013), and there-

fore are suitably aimed at processes that are com-

plex, nonlinear, time-varying, and stochastic. The

area of intelligent control is inter-disciplinary and

combines methods from other disciplines, including

modern adaptive and optimal control, learning the-

ory, fuzzy logic, and artiﬁcial intelligence (Zaknich,

2006). Neural networks, fuzzy logic and genetic algo-

rithms are the constituent techniques of the methods

in a non-classical control engineering.

Another applicable methods belong to the dy-

namic programming (DP) class. In dynamic pro-

gramming exist subclass of reinforcement learning-

based techniques such as Q-learning, R-learning, and

action-dependent heuristic dynamic programming be-

longing to model-free methods. Much of the recent

research on continuous-time reinforcement learning

has focused on model-free methods promising for the

future studies (Kamalapurkar et al., 2018).

Methods for Stated Problem. To continue exper-

iment with SABI algorithm (Zagorskis et al., 2019)

in a hybrid environment, we select appropriate Fuzzy

Intelligent control methods. One of those methods

is an ANFIS - adaptive neuro-fuzzy inference system

(Jang, 1993). Motivated by the need to accommo-

date uncertainties in system model, much of the our

research has focused on Fuzzy methods.

For our experimental setup, we started reveal-

ing the level of boredom according to Bixler and

D’Mello, arguing that there are methods how to

detect boredom during writing tasks through log-

ging the writers’ keystrokes. As authors mentioned,

keystrokes had comparably low predictive power

- roughly 11% above chance - for discriminating

engagement-neutral and boredom-neutral states. By

adding stable traits of the learners to the model, as

Bixler and D’Mello argue, prediction performance

could be notably improved.

3 HYBRID SYSTEM DESIGN AND

EXPERIMENTAL SETUP

In this section, we describe the proof of the concept

of a hybrid boredom detection system based on fuzzy

logic methods.

The system setup involves an approach that during

the learning process organized in a game-like man-

ner, learners are given small regular assignments with

a content that matches learning objectives. Learner’s

interaction with learning content is required. After-

wards, learners apply feedback to the learning envi-

ronment. The response is partially processed, ana-

lyzed, and sent to an application running machine-

learning algorithm to identify speciﬁc behavior pat-

terns. The approach is: learners use the mobile sur-

face screens with several simultaneous data channels

providing Mobile device - Host communication. In

our experimental test-bed, the host sends instructional

data (JSON format) to the application deployed and

started on a mobile client device. Similarly, data pro-

duced on a mobile device is preliminarily processed

and sent to the host for analysis and control.

In our approach (Figure 1), we organize acqui-

sition and data processing based on mobile surface

touching events logs. We measure user’s activity

time from users’ typewriting, drawing, and screen

scrolling. The proposed scheme is applicable to each

user-machine interacting experiment (a game round).

To clarify interaction model, we follow notation; we

use subscript index: i - for user-machine interaction

rounds counting.

Figure 1: Simpliﬁed user-machine interaction protocol.

One interaction round.

CSEDU 2020 - 12th International Conference on Computer Supported Education

322

In our model, a time spent by a user for the inter-

action with different learning objects (LOs) may vary

in an arbitrary time slot. During interaction we deﬁne

the user behaviour as a signal S(t) with independent

variable t - time. The signal for user k is deﬁned as

follows:

(t) = S

(t) + N

(t) (1)

Now, we clarify notation. The k

learner’s con-

scious response to assigned task S

is a process S

(t)

mixed with simultaneously upcoming noise signals

like emotional noise N

(t) and behavioural noise

(t). Equation 1 deﬁnes learner’s answer (a signal)

model responding on a learning task.

Emotional Modality detection from learner’s data

involves both measuring of attention time, spent on

LO, and reply classiﬁed as right or wrong answer in-

teracting with LOs. Since learners’ behaviour data is

encoded with a noise signal

N(t) = N

(t) + N

(t) (2)

, decoding process, in general, lead to signal S

(t)

detection errors due to noise stochastic characteris-

tics. Hence, user boredom detection problem relates

to the correct decoding of a signal transferred over a

noisy environment. Here, optimal ﬁltering techniques

can be applied.

Figure 2: Hybrid AI-Fuzzy Intelligent Control (HAFiC)

system. The model of learner’s cognitive signal S

(t)

degradation due to impact of emotional N

(t) and be-

havioural N

(t) noise. Since behavioural data contains var-

ied hidden noise sources, decoding can lead to cognitive

response data detection errors.

To continue with experimental setup description,

we clarify algorithm named as Simple Algorithm for

Boredom Identiﬁcation (SABI) recently proposed in

(Zagorskis et al., 2019). Also, we reveal crucial ex-

perimental settings structure, components, and usage

aspects before model validation experiment.

To rebuild the SABI algorithm ready for mod-

elling, authors follow the best ﬁndings and prac-

tice in the adaptive mobile learning (Cinquin et al.,

2019),(FOX, 2016). Also, into consideration are

taken some relevant ﬁndings starting from initial ideas

(Tappert et al., 1990), following through the pitfalls,

issues and challenges (Abuzaraida et al., 2013) to

recent achievements (Wang et al., 2016), that re-

searchers ﬁgured out in handwriting recognition.

In Figure 2, we illustrate components interaction

protocol for the Hybrid AI-Fuzzy Intelligent Con-

trol (HAFiC) system - the addon to online learning-

tutoring environment. During the learning process,

users interact with LOs producing a batch of vector

or raster–graphics data. AI and Fuzzy extensions to

SABI algorithm operate, allowing build sustainable

hybrid system with feedback and, what is essential,

forecasting options.

During the experiment calibration process, user

behaviour information is gathered and stored on the

host. Obtained data is also sent to the Fuzzy Expert

system. The goal of calibration is to determine spe-

ciﬁc, time-related user behaviour traits. In particular,

to identify an average (or problem-speciﬁc) response

time window, we provide experiments in game-like

approach (Figure 3). Gaming element is to control,

restrict, or just identify user response time window in

a sequence of organized experiments. Once Learner’s

Response Time increases in comparison to in calibra-

tion stored ”normal” response time - the learner is

classiﬁed as bored if his/her interaction quality addi-

tionally classiﬁed as diligent. Diligence detection is

also one of the objectives of the problem landscape.

Here, the fuzzy approach is helpful for experiment-

ing.

4 RESULTS AND DISCUSSION

In this section, we give insights into experimental data

on decision making in a fuzzy-logic equipped envi-

ronment. We represent student’s activity, which is

grouped into ﬁve classes based on recommendations

from 1) AI vector or raster–graphics data recognition

system, 2) AI learners’ diligence detection system,

and 3) explicit learners polling regarding their feel-

ings being bored (see Figure 4). We perform fuzzi-

ﬁcation as a mapping from an observed input space

to fuzzy sets in a time universe of discourse. In our

model, a time spent on a task is mapped to the enu-

meration of linguistically expressed emotional states

(bored, slow, active, fast, sharp).

Firstly, to represent the experiment data, we ap-

ply Fuzzy C-Means Clustering Algorithm to classify

First Insights into Hybrid AI-Fuzzy Tutoring System for Boredom Identiﬁcation

323

Figure 3: SABI algorithm in Hybrid AI-Fuzzy Intelligent

Control (HAFiC) system. Example work-path: educator in-

vokes N-rounds Learning Game. Calibration Round starts:

1) learner responds within user-speciﬁc time interval. 2) re-

sponse time is recorded as a ”normal” for the user. Round

One: a similar or equal task is assigned. 1) SABI algorithm

identiﬁes a gain of user-machine interaction performance -

user ﬁnishes task before calibrated time window ends. 2)

Fuzzy-Expert system algorithms compute learner’s perfor-

mance indicators and can respond with a feedback. Round

: a similar problem. 1) the user being bored - interaction

is slow or ignored. 2) SABI algorithm identiﬁes ”signif-

icant” time increase comparing to calibrated value. As a

result, 1) Fuzzy-Expert system responds to learners and ed-

ucators by giving a feedback, and 2) learner performance

model updates in HAFiC settings.

outcomes among the ﬁve subsets of the A set. The

”sharp” value corresponds to the fastest response time

and correct answer on assigned task, whereas ”bored”

state indicates the slow - non-responsive or discursive

or disjointed actions. Next, we use Gaussian Fuzzy

approximation method applied to each subset.

Finally, we depict one experiment results ( Fig-

ure 3 ). Here, on horizontal axis yellow curve indi-

cates maximum value at 1400 ms, which is the mean

value of data from all the games, and all the game

rounds classiﬁed as ”Active” since the ﬁrst - cali-

bration round. Also, it should be mentioned that in

each individual game-case ”normal” response time is

different and depends on two factors: problem com-

plexity and learner’s personality traits. Also, in pre-

sented experiment data, ”bored” class is characterized

by task completion time greater than 2 sec.

Next, on Figure 5, after data classiﬁcation and

re-scaling, we identify an unusual value of mem-

bership function associated with the emotional state

”Slow”. Such a result can be explained by an insuf-

ﬁcient amount of data - as it was already mentioned,

in our experiment (n=54). The reason for such a lim-

itation lays within the boundaries of experiment - a

limited number of involved students (54) and a one

task in an assignment.

Figure 4: Students’ activity results. Five diverse judge-

ments about user emotional states on the universe of task

completion time. Dashed line identiﬁes educator expecta-

tions regarding average task completion time. The dotted

line is the mean value of a data class ”normal activity”. In

this experiment, learners are more productive comparing to

educators forecast.

Figure 5: Students’ activity results. Five diverse judge-

ments about user emotional performance on the universe of

task completion time. Degree of membership normalized

and rescaled. The dashed line identiﬁes educator expecta-

tions regarding average task completion time.

Before the experiment, assignment outcomes also

estimated by experts’ council suggesting outcome

equal (on the average) to 1500 ms.

On another hand, during the experiment, we iden-

tify average learners’ performance improvement -

”active” rescaled curve ( on Figure 5 ) is wider and

shifted towards better performance (approx. 1200

ms).

Ultimately, comparing students’ emotional states

after each experiment round (using an explicit ques-

tionnaire method) with data classiﬁed from the output

of the experimental Fuzzy Intelligent system mockup,

we acquired the level of conﬁdence close to 70%.

CSEDU 2020 - 12th International Conference on Computer Supported Education

324

5 CONCLUSIONS

Experimental validation of SABI algorithm imple-

mentation in Hybrid AI-Fuzzy mockup system has

shown that such a hybrid system operates and is capa-

ble of solving aimed problems. Although, for build-

ing an accurate model with higher accuracy level,

there should be a sufﬁcient amount of input data.

Summing up the results, we can conclude that AI-

Fuzzy models with simple components are compar-

atively easy to implement and use but they do not

always provide good accuracy. This ﬁnding can be

explained by algorithm’s dependence on the number

of students involved in the experiment, the number of

assigned tasks on the host side, as well as the honest

answers, difﬁculty, and fuzziness, to classify boredom

by students themselves.

The accuracy of experimental model can also be

improved by the usage of complex components and

advanced system architecture but those models are

difﬁcult to implement.

An optimal choice of components for the fuzzy

model element design is especially important when

implementing the model in low-end hardware. Fur-

ther research will involve experimenting with more

data sources, wider range of tasks’ completion time

variety, ensuring feedback automatizing process and

work on the increase of data intelligence efﬁciency.

ACKNOWLEDGMENT

This research has been supported by a grant

from the European Regional Development Fund

(ERDF/ERAF) project ”Technology Enhanced

Learning E-ecosystem with Stochastic Interdepen-

dences - TELECI”, Project No.1.1.1.1/16/A/154.

REFERENCES

Abuzaraida, M. A., Zeki, A. M., and Zeki, A. M. (2013).

Problems of writing on digital surfaces in online hand-

writing recognition systems. 2013 5th International

Conference on Information and Communication Tech-

nology for the Muslim World, ICT4M 2013, pages 1–5.

Bixler, R. and D’Mello, S. (2013). Detecting boredom and

engagement during writing with keystroke analysis,

task appraisals, and stable traits. In Proceedings of

the 2013 international conference on Intelligent user

interfaces - IUI ’13, page 225, New York, New York,

USA. ACM Press.

Bradley, M. M. and Lang, P. J. (1994). Measuring emotion:

the self-assessment manikin and the semantic differ-

ential. Journal of behavior therapy and experimental

psychiatry, 25 1:49–59.

Cannarile, F., Compare, M., Baraldi, P., Di Maio, F.,

and Zio, E. (2018). Homogeneous Continuous-Time,

Finite-State Hidden Semi-Markov Modeling for En-

hancing Empirical Classiﬁcation System Diagnostics

of Industrial Components. Machines, 6(3):34.

Cinquin, P.-A., Guitton, P., and Sauz

eon, H. (2019). Online

e-learning and cognitive disabilities: A systematic re-

view. Computers & Education, 130:152–167.

FOX, A. M. Y. (2016). Microlearning for Effective Perfor-

mance Management. (cover story). TD: Talent Devel-

opment.

Graf, S., Liu, T.-C., and Kinshuk (2010). Analysis of learn-

ers’ navigational behaviour and their learning styles in

an online course. J. Comp. Assisted Learning, 26:116–

131.

Jang, J. . R. (1993). Anﬁs: adaptive-network-based fuzzy

inference system. IEEE Transactions on Systems,

Man, and Cybernetics, 23(3):665–685.

Kamalapurkar, R., Walters, P., Rosenfeld, J., and Dixon, W.

(2018). Reinforcement Learning for Optimal Feed-

back Control: A Lyapunov-Based Approach. Com-

munications and Control Engineering. Springer Inter-

national Publishing.

Kwiatkowska, M. Z. (2013). Advances in quantitative veri-

ﬁcation for ubiquitous computing. In ICTAC.

Li, L. X. and Rahman, S. S. A. (2018). Students’ learning

style detection using tree augmented naive bayes. In

Royal Society open science.

Lugo, G. S., Rodr

ıguez, L.-F., L

opez, R. I. G., Mac

ıas-

Estrada, A., and Echeverr

ıa, M. R. (2015). Behavioral

patterns for automatic detection of learning styles in

learning management systems: a case study. Research

in Computing Science, 106:69–77.

Pielot, M., Dingler, T., Pedro, J. S., and Oliver, N. (2015).

When attention is not scarce - detecting boredom from

mobile phone usage. In Proceedings of the 2015

ACM International Joint Conference on Pervasive and

Ubiquitous Computing, pages 825–836, Osaka. ACM.

Russell, S. and Norvig, P. (2016). Artiﬁcial Intelligence: A

Modern Approach. Always learning. Pearson.

Siddique, N. (2013). Intelligent Control: A Hybrid Ap-

proach Based on Fuzzy Logic, Neural Networks and

Genetic Algorithms. Studies in Computational Intelli-

gence. Springer International Publishing.

Tappert, C. C., Suen, C. Y., and Wakahara, T. (1990). The

State of the Art in On-Line Handwriting Recognition.

IEEE Transactions on Pattern Analysis and Machine

Intelligence.

Tze, V. M. C., Daniels, L. M., and Klassen, R. M. (2016).

Evaluating the relationship between boredom and aca-

demic outcomes: A meta-analysis. Educational Psy-

chology Review, 28:119–144.

Wang, G., Zhang, X., Tang, S., Zheng, H., and Zhao, B. Y.

(2016). Unsupervised Clickstream Clustering for User

Behavior Analysis. In Proceedings of the 2016 CHI

Conference on Human Factors in Computing Systems

- CHI ’16.

Wang, N., Sun, S. D., Cai, Z. Q., Zhang, S., and Saygin, C.

(2014). A hidden semi-markov model with duration-

First Insights into Hybrid AI-Fuzzy Tutoring System for Boredom Identiﬁcation

325

dependent state transition probabilities for prognos-

tics. Mathematical Problems in Engineering, 2014(1).

Zagorskis, V., Kapenieks, A., and Gorbunovs, A. (2019).

Emotions identiﬁcation utilizing periodic handwriting

on mobile surfaces. Periodicals of Engineering and

Natural Sciences (PEN), 7(1):IC18.

Zaknich, A. (2006). Principles of Adaptive Filters and Self-

learning Systems. Advanced Textbooks in Control and

Signal Processing. Springer London.

CSEDU 2020 - 12th International Conference on Computer Supported Education

326